Time-Series Clustering and Visualization for Insights into Multimorbidity Progression

The increasing availability of vast amounts of data in electronic health records (EHR) offers immense opportunities to extract valuable insights, particularly through the application of machine learning techniques like clustering. This thesis focuses on clustering time-series data extracted from medical records, with the aim of identifying meaningful clusters of patient sequences. While clustering methods are well-established for static datasets, clustering time series data presents unique challenges, especially when it comes to selecting the most relevant solution from many valid clustering outputs.
In this work, we develop a two-stage methodology for clustering time-series data. The first stage simplifies high-dimensional sequence data, while the second stage focuses on identifying clusters within these sequences. We also address the issue of comparing multiple clustering solutions by introducing a novel approach that combines a graphical user interface (GUI) with a graph-based representation of the relationships between different clustering solutions. This framework allows for intuitive, simultaneous exploration and comparison of multiple valid solutions, helping to reduce the space of possible results and aiding in the interpretation of alternative outcomes.
Our methodology is applied to the domain of multimorbidity, a significant healthcare
challenge characterised by the coexistence of multiple chronic conditions. By applying our tools to multimorbidity datasets, we gain insights into the progression of chronic illnesses and their interactions.